README.md

---
title: NumidiaNLP
emoji: 🌍
colorFrom: red
colorTo: yellow
sdk: static
pinned: false
---

<div align="center">

# NumidiaNLP

### Natural Language Processing for Maghrebi Arabic

**Language Models · Resources · Evaluation · Reproducible Research**

</div>

---

## About

**NumidiaNLP** is a research initiative dedicated to Natural Language
Processing for **Maghrebi Arabic and North African digital language**.

Our research focuses on the linguistic diversity observed across
**Algeria, Morocco, and Tunisia**, including dialectal Arabic,
non-standard orthography, **Arabizi**, and multilingual
**code-switching**.

The broader objective is to develop reusable language models, datasets,
and evaluation resources for underrepresented Maghrebi language varieties.

---

## Research Focus

- Maghrebi Arabic language modeling
- Regional continued pretraining
- Algerian, Moroccan, and Tunisian Arabic
- Arabizi and Latin-script Maghrebi Arabic
- Arabic–French and multilingual code-switching
- Hate and offensive language detection
- Sentiment analysis
- Dialect identification
- Corpus construction and annotation
- Robust and reproducible NLP evaluation

---

## 🌍 MagharibiBERT

### A Regionally Adapted Language Model for Maghrebi Arabic

**MagharibiBERT** is a regional language model developed to provide
reusable contextual representations for heterogeneous Maghrebi Arabic.

It is adapted through large-scale continued pretraining over diverse
Maghrebi text representing linguistic variation from Algeria, Morocco,
and Tunisia.

The training resources preserve important characteristics of real-world
Maghrebi communication, including dialectal variation, non-standard
orthography, Arabizi, and multilingual code-switching.

The central research question is:

> **Can regional continued pretraining provide transferable language
> representations across Maghrebi countries and NLP tasks?**

🤗 **Model:**  
https://huggingface.co/NumidiaNLP/MagharibiBERT

---

## Regional Perspective

Maghrebi Arabic is not treated as a single homogeneous dialect.

Instead, we consider it a **linguistically diverse regional continuum**
with substantial variation across countries, communities, scripts,
platforms, and communication contexts.

Our research therefore investigates both:

- the value of **shared regional representations**, and
- the complementary strengths of **country-specific models**.

---

## Evaluation Principles

Our experimental methodology emphasizes:

- fixed train/validation/test partitions;
- multiple random seeds;
- validation-based checkpoint selection;
- Macro-F1 for imbalanced classification tasks;
- comparisons with general Arabic and regional language models;
- statistical analysis where appropriate;
- reporting of computational requirements;
- transparent reporting of limitations and negative results.

The objective is not merely to obtain the highest score on one dataset,
but to study **cross-country robustness, stability, and transferability**.

---

## Resources

| Resource | Description |
|---|---|
| 🤗 **MagharibiBERT** | Regional language model for Maghrebi Arabic |
| 📚 **Models** | Maghrebi and task-adapted NLP models |
| 📊 **Benchmarks** | Cross-country and cross-task evaluation |
| 🗂️ **Datasets** | Documented Maghrebi NLP resources when licensing permits |

---

## Reproducible Research

Where publication, licensing, and ethical requirements permit,
NumidiaNLP aims to release:

- pretrained models;
- evaluation code;
- benchmark configurations;
- dataset documentation;
- reproducible experimental protocols.

Each released resource is accompanied by documentation describing its
provenance, intended use, evaluation setting, and limitations.

---

<div align="center">

### NumidiaNLP

**Advancing NLP research for the linguistic diversity of the Maghreb**

🇩🇿 Algeria · 🇲🇦 Morocco · 🇹🇳 Tunisia

</div>